COBALT detects tuberculosis from cough audio with new fusion framework

A team of researchers led by Mohd Mujtaba Akhtar have developed COBALT, a novel framework for non-invasive tuberculosis (TB) detection from cough audio. The framework fuses two types of representations: spectral descriptors like MFCCs (mel-frequency cepstral coefficients) which capture fine-grained acoustic details, and foundation embeddings like PaSST (a pretrained audio spectrogram transformer) which encode higher-level temporal patterns. To integrate these heterogeneous features, COBALT uses codebook-aligned hyperbolic prototypes and a bandit-style reliability weighting mechanism, which dynamically weights each representation based on its contribution to the screening task.

Tested on the CODA TB DREAM Challenge benchmark—a standard dataset for cough-based TB screening—COBALT consistently outperformed individual representations and a simple concatenation baseline. The best performance was achieved by fusing MFCC and PaSST, establishing a new state-of-the-art on the benchmark. The paper has been accepted to INTERSPEECH 2026. This work demonstrates the potential of combining classical audio features with modern deep learning embeddings for medical diagnostics, paving the way for low-cost, scalable TB screening tools that could be deployed in resource-limited settings.